Chapter 10. Awareness driven communication
10.3.16. Examples
showing the improvement throughout. For example, if abstractions were sufficient for 75% of artefacts, and abstraction resulted in a 10-fold reduction in resource requirements on average (i.e. = 0.75 and = 0.1) then the single partici- pant bandwidth would be reduced by 67% while the total network bandwidth would, for the three scenarios in table 14 on page 144, change by +9.1%, -30.5% and -61.8%, respectively (for the teleconference, visualisation and race). Larger values of and smaller values of would produce larger savings.
It is apparent that secondary sourcing and abstraction have the potential to signifi- cantly reduce the load on a single participant if abstracted views of many artefacts are acceptable to the user. Whether this is the case will depend on the application. In many scenarios the use of abstractions may also significantly reduce the total band- width requirements because of the reduced demand for state transfers.
10.3.16. Examples
This section concludes this analysis by applying the network traffic model to the MASSIVE-1 and MASSIVE-2 systems described in this thesis. These are considered in turn.
MASSIVE-1
MASSIVE-1 replicates based on artefacts and uses explicit state transfers rather than DIS-style heartbeat messages. The following system-specific parameters apply for MASSIVE-1 based on the analysis in chapter 6: = 13.2 Kbytes, = 0, = 4.6 Kbytes/s, = 2100 bytes and = 1400 bytes/s (approximately, including inter-client communication). Using artefact-based replication (and assuming the suffi- ciency of spatial trading to represent scope of interest) = 1, = 1 and = . Networking is unicast and so total network bandwidth (from equation 7) is:
(Equation 10-15) Similarly the per-participant bandwidth (from equation 6) is:
(Equation 10-16) Chapter 6 also neglected and took to be per second, giving a total network bandwidth of bytes/s (c.f. equation 1 on page 73, , where and ). For this case the single par- ticipant bandwidth is bytes/s.
1–fI(1–fB) ( ) fI fB fI fB S BA BP SRI BRI EM AI NRI IA+IP 15400 (IA+2IP) M TS --- ⋅ +4600 I⋅ P+1400 N P 15400 (IA+2IP) M TS --- ⋅ +4600 I⋅ P+1400 IA M T⁄ S 1 60⁄ 4850IP+1400 ( )NP B = N 4800M( +1400) M = IP N = NP 4850IP+1400
10.4. Summary and conclusions
MASSIVE-2
MASSIVE-2 performs replication based on worlds and regions, uses explicit state transfers like MASSIVE-1 and uses multicast communication for updates. The fol- lowing system-specific parameters apply for MASSIVE-2 (see section 10.3.3): = 1628 bytes, = 0, = 2495 bytes/s, = 1624 bytes and = 0. The total net- work bandwidth will be:
(Equation 10-17) The single participant bandwidth will be:
(Equation 10-18) For square cells as analysed in section 10.3.12 (and table 16 on page 148):
,
For example, consider = , = 0 (as for MASSIVE-1, above) and = 8.45, and (comparable to NPSNET cells as in section 10.3.12). This gives a total network bandwidth of bytes/s. This is a 38-fold reduction compared to MASSIVE-1 for large values of . For = 10 (as in the ITW trials) the reduction in bandwidth requirements is by a factor of 11. The single participant bandwidth is . So the reduced accuracy of open cells results in a typically higher average bandwidth for each participant than in MASSIVE-1 in this type of environment.
This reduced accuracy and correspondingly increased per-participant bandwidth is a general consequence of using a larger unit of replication (regions rather than arte- facts). This choice is in turn motivated by the adoption of multicast communication, with system and network restrictions and overheads linked to the number of multicast groups employed.
However when an environment includes closed regions then MASSIVE-2’s exploita- tion of third party object effects would make it more accurate than MASSIVE-1’s aura-only approach (and other approaches which assume a nominally open space). This would cause a corresponding reduction in participant bandwidth.
10.4. Summary and conclusions
This chapter has described how the spatial model of interaction and third party objects are used in MASSIVE-2 to manage artefact replication and multicast groups.
By using an explicit computational model of awareness with third party effects the system is able to represent a range of situations from open terrain to closed rooms
S BA BP SRI BRI 1628 IA+IP 1 EM --- + 1624N I + M TSAI ---+2495 N P 1628 IA+IP IP EM --- + 1624NI + M TSAI --- 2495IP AI --- + AI πNI 4 NI+12 --- = EM 1 NI --- = M T⁄ S 1 60⁄ IA NI AI = 0.435 EM = 0.344 125IP+3100 ( )NP IP Ip 5877IP
10.4. Summary and conclusions
with reasonable accuracy. This is achieved through three forms of replication man- agement: aura-based, membership-based and awareness-based. All three are demon- strated in the new audio gallery world (described in section 10.2).
The majority of this chapter has developed a network bandwidth model for CVEs based on experience gained over the course of the work presented in this thesis. This traffic model is used to analyse a number of distribution and communication issues in CVEs and similar systems. The results are summarised below.
• State and updates. The model shows that the relative requirements of state transfers compared to (multicastable) updates vary widely for different application scenar- ios. In extreme applications (such as the “race” example in table 14 on page 144) state transfers can dominate, especially when multicasting is used for updates. In situations such as this multicast state transfers or group aggregates might be used to reduce the total network bandwidth requirements (though multicast state trans- fers will not affect the participant bandwidth).
• Heartbeat-based state transfers. For applications which involve many mutually aware participants (i.e. large ) and which are limited by total network bandwidth rather than participant bandwidth then using heartbeat messages to perform state transfer can reduce total network bandwidth requirements. However this will typi- cally result in a significant increase in single participant bandwidth, as well as the need to trade off timeliness of information against bandwidth requirements through the choice of , the heartbeat time. In general explicit state transfers are more appropriate. More specific (reliable) multicast state transfers or abstractions could be used if state bandwidth requirements remained the limiting factor for scalability. • Multicast and unicast. The benefits of using multicasting in terms of total network bandwidth can be very great, especially if the update bandwidth is relatively high (e.g. with video) or if some form of multicasting is also be employed for state transfers. The use of multicasting reduces the number of messages which a partici- pant has to send but it does not reduce the number which they will receive. In fact the use of multicasting may lead to the use of a less accurate form of replication management (see below) so that each participant actually has to deal with more information. However, except on very high-speed LANs (e.g. workgroup ATM) total network bandwidth limitations will mean that multicasting will be a necessary choice for significant numbers of mutually aware participants.
• Replication unit. Replication can be performed at a number of levels of granularity, classified here as “universe”, “world”, “region” and “artefact”. The universal repli- cation approach is unnecessarily limited. The world-based approach is simple and useful, but depends on being able to organise interest and interaction to match. The region approach is more flexible than the world approach as demonstrated in MASSIVE-2. Finally the artefact approach is potentially the most accurate (a par- ticipant only replicates what they need) but the management overheads will typi- cally be greatest as will the network resource requirements (e.g. the number of multicast groups or reservations required). The best compromise might be to com- bine general region-based replication with artefact-based replication for extremely demanding artefacts (e.g. which include real-time video streams).
• Replication management. Depending on the choice of replication unit replication management can be performed in a number of ways. The goal is to replicate only those artefacts which are of direct interest to a given participant. This requires a
IP
10.4. Summary and conclusions
flexible and expressive way of representing the participant’s interest and an accu- rate way of mapping this onto units of replication. This is one strength of MASSIVE-2 and the spatial model of interaction with third party objects. More limited forms of replication management (e.g. using open cells or based solely on occlusion) will be either less accurate or more limited in the range of applications which they can address.
• Secondary sourcing and abstraction. Secondary sourcing and abstraction have the potential to significantly reduce the load on a single participant if abstracted views of many artefacts are acceptable to the user; this will depend on the application. Also, where unicast state transfers are the dominant component of total network bandwidth then the use of abstractions may also significantly reduce total band- width requirements.
The model and methodology used in this analysis could also be applied and general- ised to other situations and to other classes of application. The notion of awareness (which need not be explicit in the application) provides the essential basis for the model and allows reasoning about, for example, accuracy, which is one of the key components and outcomes of the model.
This concludes part II of this thesis which has considered MASSIVE-2 and the third party object concept. The next and final chapter draws together the various themes running through this thesis and presents the final conclusions and suggestions for future work.
10.4. Summary and conclusions
Table 18: summary of network model parameters
Parameter Meaning
The accuracy of replication compared to a participant’s ideal scope of interest.
The bandwidth (in bytes per second) generated by a participant. The bandwidth (in bytes per second) generated by a passive artefact (often zero).
The continuous bandwidth (bytes per second) associated with repli- cation management generated by each participant.
The “efficiency” with which mobility is handled.
The number of passive artefacts which fall within an average scope of interest.
The number of participants which fall within an average scope of interest.
Mobility of interest, i.e. the number of times during a session that the participant moves such that the artefacts and participants within their scope of interest are completely replaced by new ones.
Number of artefacts (passive) spread over all worlds.
Number of simultaneous participants spread over all worlds. The total number of replication units used by the system.
The average number of replication units of interest to a single partic- ipant at any one time (ideally).
The total number of worlds over which artefacts and participants are distributed
The size (in bytes) of the state of an average artefact or participant’s embodiment.
The network traffic (bytes) associated with each paging of a replica- tion group (in and out, total).
The length of time for which a participant uses the system in a sin- gle session. AI BP BA BRI EM IA IP M NA NP NR NRI NW S SRI Ts